In the early stages of child development, identifying and intervening in developmental delays or psychological issues is crucial. Traditional methods rely on observations by parents and educators, as well as professional assessments, which are limited by their subjectivity, difficulty in quantification, and lack of continuous monitoring data, leading to diagnostic delays or misdiagnoses. To address these issues, we propose a new model combining deep learning with biological signal analysis— MRANet—for the early detection of abnormal behaviors in children. The model consists of three core modules: the Multimodal Representation Module (MMR), the Riemannian Feature Mapping Module (RFM), and the Multi-Source Interactive KAN Module (MSI-KAN). First, biological signals such as EEG and ECG are collected through simple sensors, segmented into 1-s time blocks, and processed through a Butterworth filter to transform them into three-dimensional representations across four frequency bands: Delta, Theta, Alpha, and Beta. Next, in the RFM module, the spatial covariance of physiological signals is extracted as features, and a bilinear mapping layer is used to adjust the symmetric positive definite matrix (SPD) distribution. The Log-Euclidean metric is then applied to reduce these features from the non-Euclidean manifold to the flat space. Finally, the MSI-KAN module receives data from both 1D signals and 3D image modalities, performing nonlinear feature extraction through the KAN network. The randomly concatenated features are then fed into the classifier to obtain the final prediction results. Experimental results show that MRANet achieves accuracies of 94.7% and 92.5% in binary and five-class classification tasks, respectively, significantly outperforming existing methods and demonstrating its tremendous potential in safeguarding children’s mental health.